Method for supporting medical procedures and implant design using fluid flow simulation and machine learning

Abstract

A method for supporting a medical procedure, the method comprising: obtaining a three-dimensional image of a vessel of a patient; deriving measurements of the vessel from the three-dimensional image; constructing a model of fluid flow in the vessel using the measurements; and using a trained machine learning model to make predictions about the medical procedure based on outputs of the model of fluid flow. The medical procedure may comprise implantation of a device and the predictions may comprise at least one of: a recommendation of a size or type of device; a recommendation of a placement of the device; and a predicted lifetime of the device in situ. The medical procedure may comprise an Aortic Valve Replacement, desirably a Transcatheter Aortic Valve Replacement or a valve-in-valve procedure.

Description

MEDICAL SUPPORT METHODField of the Invention

[0001] The present invention relates to a medical support method involving characterisation of body passages, e.g. vessels or valves of the cardiovascular system, based on imaging and in particular determination of characteristics of body vessels relevant to implantation of medical devices, e.g. replacement valves.Background

[0002] Aortic stenosis is a condition characterised by the restriction or narrowing of the aortic valve. As a consequence, the valve doesn’t open fully, restricting blood flow from the left ventricle into the aorta and the rest of the body. As a result, the heart must work harder to pump blood through the valve, potentially leading to symptoms such as chest pain, fainting, shortness of breath, and heart failure if left untreated. Most aortic stenosis cases are caused by calcification of the aortic valve leading to valve degeneration. The number of calcified aortic valve disease cases worldwide has been estimated at 9.4-12.6 million people.

[0003] A common treatment for aortic stenosis is replacement of the defective valve. Surgical aortic valve replacement (SAVR) is an open chest surgery where an incision is made to access the heart, which is stopped and a heart-lung by-pass machine is used to take over the heart’s function during the operation. The valve is replaced, heart restarted and chest surgically closed.

[0004] Alternatively, Transcatheter Aortic Valve Replacement (TAVR) offers a minimally invasive procedure. TAVR involves the placement of a new prosthetic valve into the heart via a catheter, which is inserted through a small incision (for example, in the leg) and guided to the heart. This new valve is deployed within the existing aortic valve, where it expands and takes over the function of the diseased valve, thereby restoring normal blood flow into the aorta. Early clinical studies focused on the application of TAVR to high-risk patients, but indications for TAVR have expanded and now include younger and low and intermediate-risk patients.

[0005] Several companies manufacture and supply valves for TAVR interventions, contributing to a diverse market of devices designed to accommodate the varied anatomical and clinical needs of patients. Within a single company’s product line, there are multiplevalve models available, each with distinct properties, including various sizes and design features. These differences allow for a personalised approach to TAVR, enabling the patient’s multidisciplinary heart care team to select the most suitable valve based on specific factors such as the patient’s aortic annulus size and other considerations. The location and orientation of the implanted valve directly influence hemodynamics by affecting the flow dynamics and pressure gradients across the valve, thereby determining the efficiency of blood ejection from the heart into the systemic circulation. Flow dynamics and vorticities of blood within the aorta, influenced by the position and orientation of the implanted valve, impact aortic wall shear stress and may affect the structural integrity and durability of the implanted valve. The success of a TAVR intervention depends on the correct choice of valve to match the patient’s anatomy, along with its positioning and orientation to optimise the hemodynamics after implantation and the longevity of the valve.

[0006] Computational modeling in TAVR represents a potentially useful tool for clinicians, enhancing understanding of post-procedural performance and potential complications, thereby guiding device design, procedural planning, and patient-specific care. The integration of patient-specific data into simulations helps in defining models that reflect individual anatomical and physiological characteristics, offering insights to optimize procedural outcomes.

[0007] Clinical considerations such as thrombosis, paravalular leak (PVL) and the durability of the TAVR device are important considerations when planning a TAVR intervention. Khodaei et al.

[0010] , Singh-Gryzbon et al. [1] and Hatoum et al. [5] provide patient-specific computational modeling to reveal the hemodynamics resulting in thrombus formation. These models emphasize the importance of considering individual patient characteristics, like the anatomy of the aortic root and aortic valve, to improve TAVR strategies and potentially reduce thrombotic risks. Paravalvular leak has been studied using computational methods such as finite element analysis (FEA) and computational fluid dynamics (CFD) to inform deployment and positioning of the device. For example, Bianchi et al. [2] used FEA and fluid-structure interaction (FSI) to understand the effects of valve positioning on PVL outcomes, underscoring the significance of device interaction with the patient’s unique anatomical geometry. Patient-specific computational modeling has become pivotal in assessing TAVR device durability, offering insights into the nuanced interactions between biomechanical forces and bioprosthetic valve longevity. Fumagalli et al. [4] used FSImodels to study the impact of systolic wall shear stress in predicting structural valve degeneration (SVD).

[0008] Pre-interventional planning can assist the multidisciplinary heart team optimise TAVR outcomes, and digital tools have been developed for this purpose. Toggweiler et al.

[0018] developed a cloud-based Al-driven software called 4TAVR that enables fully automated TAVR CT analysis. Based on deep learning, the method performs segmentation of the cardiac structures, and aortic root anatomical landmarks are identified on the aortic annulus, sinus and ostia along with a centreline. 4TAVR outputs a number of key measurements in a PDF report used by clinicians for sizing. Weber et al.

[0019] present a segmentation approach that helps in TAVI implant positioning using transesophageal echocardiography.

[0009] Rouhollahi et al.

[0016] propose a fully-automated deep learning approach called CardioVision that creates a digital replica of a patient’s aortic root, native valve, and calcification from a patient’s pre-interventional CT scan. Geometries of the aorta and calcifications are produced in the form of 3D models in STL format, suitable for subsequent visualization and digital twin simulation of interventional procedures.

[0010] Kadry et al. [7] propose the use of Generative Al in the form of diffusion methods to create anatomic variants of digital cardiac twins, termed digital siblings. These models provide virtual cohorts for device assessment and in silico trials.

[0011] In terms of simulation, Kuchumov et al.

[0011] provide a review article current as of October 2023 summarising work performed on fluid-structure interaction involving computational fluid dynamics and finite element methods for predicting outcomes of aortic valve replacement. The paper highlights the value of computational modeling but also discusses the challenges around simulation instabilities, computational complexity, lack of medical data and the importance of turbulence amongst other findings. Relatedly, Tahir et al.

[0017] provide a summary of deep learning approaches to assessing TAVR procedures and outcomes current as of July 2023.

[0012] Kandail et al. [8] simulate hemodynamics using fluid-structure interaction for a Medtronic CoreValve deployed in annular and supra-annular locations to study the blood velocity and wall sheer stress (WSS). The study also explores the effect of valve orientation, finding a supra-annularly deployed valve with a lateral tilt of 10 degrees led to a more centred jet. Dowling et al. [3] also perform patient-specific simulation to optimise valve sizing and positioning to minimise the risk of paravalvular regurgitation (PVR).

[0013] Pietrasant et al.

[0015] study the effect of turbulence after a transcatheter aortic valve implanted into a phantom mimicking the aortic root. The paper found the degree of device expansion in situ relates to the onset of turbulence, and a smaller and less regular opening area can introduce flow instabilities that may be detrimental for the long-term performance of the valve.

[0014] Given a medical image such as a CT image, a number of processing steps such as segmentation and meshing must be performed prior to blood flow simulation. DeepCarve

[0014] and its extension C-MAC

[0013] provide a deformation-based approach to fast mesh creation from a CT image. To be simulation ready, meshes must have smooth and clearly defined ends with boundary surfaces clearly defined. Open source solutions such as the vascular modelling toolkit (VMTK www.vmtk.org / ) can provide segmentation, centrelines, and mesh generation, and this tool has been ported to 3D Slicer (github.com / vmtk / SlicerExtension-VMTK).

[0015] There are several open source software solutions for computational fluid dynamics designed for cardiovascular anatomies. These include Crimson(crimson. software / index.html) and SimVascular (simvascular.github.io / ), each capable of simulating for patient-specific blood flow. However, neither have purpose-built TAVR pipelines. In practice there is often tedious manual work required to manipulate a segmentation and perform clipping prior to simulation.

[0016] Commercial software for TAVR planning includes Dasi Simulations (dasisim.ai / ) which produces FDA-approved software called Precision TAVI that creates a digital twin of a patient’s heart from a CT angiogram and simulates the heart’s interaction with a TAVI device. FEOps (www.feops.com / product / healthcare-professionals) has software called HEARTguide that lets physicians plan TAVI procedures to predict which patients are at risk for new pacemakers or paravalvular leaks via aortic root assessment and patient-specific digital simulations. However, neither Dasi Simulations or FEOps solutions provide recommendations on which device to implant, where to implant it, or predicted success metrics of the intervention.Summary

[0017] There is a need for methods and systems to support clinicians in planning interventions in a vessel of a patient, in particular involving implantation of devices, by determining relevant characteristics of the vessel.

[0018] According to an aspect of the invention, there is provided a method for supporting a medical procedure, the method comprising: obtaining a three-dimensional image of a vessel of a patient; deriving measurements of the vessel from the three-dimensional image; constructing a model of fluid flow in the vessel using the measurements; and using a trained machine learning model to make predictions about the medical procedure based on outputs of the model of fluid flow.

[0019] Desirably, the medical procedure comprises implantation of a device and the predictions comprise at least one of: a recommendation of a size or type of device; a recommendation of a placement of the device (e.g. height and orientation); and a predicted lifetime of the device in situ.

[0020] Desirably, the medical procedure comprises an Aortic Valve Replacement, desirably a Transcatheter Aortic Valve Replacement, a TAVR valve-in-valve procedure or a valve-in-valve procedure.

[0021] Therefore, embodiments of the invention can assist in planning an intervention, in particular by providing information about a specific patient and operation site, for example relating to potential flow dynamics and its impact on wall shear stress, turbulence and valve durability, relevant to decisions such as valve selection, its positioning based on the aortic root geometry, and orientation.

[0022] Embodiments of the present invention may provide a model of an operation site calibrated to a specific individual that is referred to herein as a “digital twin”. Such a digital twin can offer a transformative approach to TAVR pre-interventional planning by enabling detailed modeling and simulation of an individual patient’s heart anatomy and function. By incorporating patient-specific data, such as anatomical details from imaging studies and physiological parameters, digital twins can provide simulations for the placement of different valve types and sizes, their locations, and orientations within the heart. These advanced simulations allow for the analysis of potential hemodynamic outcomes, including flow dynamics, vorticities, and wall shear stress in the aorta, particularly in the aortic root where the device is placed, providing invaluable insights into how these factors might influence the success of the intervention and the longevity of the implanted valve. The present invention can enable an improved haemodynamic outcome - i.e. the maximal reduction in the pressure gradient across the aortic valve and the absence of aortic regurgitation / paravalvular leak.

[0023] Embodiments of the present invention provide a digital twin pipeline that provides mechanistic simulation of a patient’s blood flow coupled with predictive capabilities of artificial intelligence. TAVR-AID is designed as a pre-interventional decision-support tool, helping the heart care team responsible for the TAVR procedure anticipate and mitigate potential complications, optimise valve selection and placement, and tailor the TAVR procedure to the unique needs of each patient, thereby enhancing patient care.Brief Description of the Drawings

[0024] The present invention will be described below with reference to exemplary embodiments and the accompanying drawings, in which:Figure 1 is a schematic diagram of a method of an embodiment of the invention;Figure 2 depicts example CT image segmentations;Figure 3 provides examples of clipping;Figure 4 provides examples of meshing steps;Figure 5 depicts modeling the inflow boundary condition; andFigure 6 depicts simulation results.

[0025] In the various figures, like parts are denoted by like references.Detailed Description

[0026] An embodiment of the invention is a digital twin pipeline for transcatheter aortic valve replacement. The digital twin provides patient-specific simulation and machine learning prediction performed prior to TAVR intervention, to provide clinical decision support to the multi-disciplinary heart team in planning the intervention to optimise the valve’s operation and longevity once implanted. The tool is not designed to replace human decision making, rather it is intended to assist human planning to produce a better outcome than if the digital twin had not been used.

[0027] It will be appreciated that the principles of the invention may also be applied to other vessels and other implanted devices. For example, the invention may be applied to the mitral valve or the tricuspid valve. For example, the invention may be applied to the implantation of stents, to provide recommendations of placement and predictions of long-term outcomes such as restenosis. In particular, the invention may be applied to: coronary stents(implanted in coronary arteries); carotid artery stents (carotid arteries); peripheral vascular stents (peripheral arteries: arms, legs).

[0028] Figure 1 provides a summary of the pre-interventional digital twin TAVR pipeline. The pipeline begins with medical images 11 of the patient. An aortic computed tomography (CT) scan provides a structural view of the patient’s 3D aortic anatomy. This image can be a CT angiography image, or a best systole or diastole time point in the cardiac cycle. Quality control is applied to ensure the image has appropriate definition of the vascular anatomy including the left ventricular outflow tract (LVOT). A segmentation algorithm 12 (e.g. nnUNet [6]) is applied to the CT image to produce a personalised 3D model representing the patient’s aortic anatomy. In this embodiment, Al is used primarily in the segmentation stage, and the final Al module at the end of the pipeline; however, the entire pipeline can benefit from Al automation. As an alternative to the CT image, it is possible to use other 3D modalities such as MRI or 3D ultrasound to build the model.

[0029] In addition to the CT image, pre-interventional echocardiography 16 is performed to assess the hemodynamic severity of the aortic stenosis and capture the left ventricular outflow peak velocity in addition to other measurements such as the diameter of the LVOT and aortic valve area (AV A). The echocardiography data is used to provide boundary conditions 17 for subsequent blood flow modeling using computational fluid dynamics.

[0030] Next, the segmentation is refined 14 to prepare a 3D mesh 15 suitable for computational fluid dynamics. The CT segmentation includes the aortic root and ascending aorta, it may also include additional regions that are not required for the simulation. Therefore, a clipping operation is performed to remove parts of the anatomy that are not required for the digital twin model. The mesh is clipped at the correct anatomical location matching the echocardiography inflow into the aorta. The computational fluid dynamics simulation also requires flat, closed, meshable ends normal to the aorta. The operations can be performed manually using software tools such as 3DSlicer (www.slicer.org) and SimVascular (simvascular.github.io / ), or performed automatically using Al and computational geometry methods. Next, the data is transformed, using a meshing operation (meshtools program) producing a “simulation grade” solid model 18.

[0031] The aortic valve is simulated at the aortic annulus, and a jet of fluid is modeled based on the echocardiographic data, providing a personalised boundary condition for blood flow into the aorta. The inflow into the aorta is modelled as a circular jet, with a radiusmatching the patient’s aortic valve annual and flow velocity from the echocardiography scan. Additional boundary conditions include assumptions of a rigid, non-permeable aortic wall where the velocity is zero, and outflow conditions such as backflow stabilisation or OD vasculature models.

[0032] A solver 19 (such as using CARPentry-Pro’s FluidSolve, although others are possible, such as OpenFOAM, Ansys or Abaqus) is then run to simulate the blood flow from the valve into the aorta, captured with high spatial resolution, using computational fluid dynamics. The simulation results 20 provide 3D + time flow patterns in the aorta. These flow patterns are then analysed by artificial intelligence looking for signatures in the hemodynamics as well as characterisation of velocities, turbulence, circulation, and stagnation particularly in the aortic root along with stresses on the aortic wall. Turbulence can be measured using Q-criterion, vorticity magnitude, or other measures. Two approaches are possible. In the first approach, intermediate values (sometimes referred to as “handcrafted features” albeit they are generated automatically) are derived from the 3D+time flow. These can include one or more of: the distribution of the velocity of the blood, number of vorticities, Reynold’s number relative to the valve, peak wall shear stress and other turbulence-related flow quantities. In a second approach, the 3D+time flow simulation is input to a deep network which would learn the features that best predict the outputs without needing to generate the intermediate values.

[0033] Based on the flow patterns 20, patient information 22 contained in the electronic healthcare record, and measurements from the CT and echo images, the artificial intelligence model 21 makes recommendations for the intervention, e.g. including one or more of: which valve to use, the valve’s size, its positioning (height and orientation), and risk of complication such as thrombosis, stroke, paravalvular leak or device failure. For the last point, the Al model can predict the longevity of the device based on the simulated blood flow.

[0034] The pipeline described above is one possible embodiment of the TAVR digital twin pipeline. More sophisticated versions of the pipeline may include a simulation of the valve. Fluid- structure interaction methods between the simulated blood flow and the leaflets of the valve show how the valve affects the blood flow, and the degree to which the leaflets open based on the position and orientation of the valve. As in the embodiment illustrated in Figure 1, hemodynamics as well as characterisation of velocities, turbulence, circulation, and stagnation, particularly in the aortic root along with stresses on the aortic wall are computed,the simulated blood flow is input into an artificial intelligence module which makes recommendations on the valve type, size, position, orientation. It can also predict the risk of complications and the longevity of the valve. In an alternative embodiment, the valve can be virtually inserted into the 3D aortic mesh. The deformation of the implanted valve at a desired implantation depth and orientation can be predicted using machine learning methods based on observed interactions between the cylindrical prosthetic valve and the native aortic valve in the training data. Paravalvular leak can be estimated based on the fit of the valve inside the aortic annulus, and the deformed shape used for fluid-structure interaction simulation. Simulating the valve in this way can be used to inform valve design, as will be discussed later.

[0035] The TAVR-AID pipeline includes the specification of appropriate input images and flow data, steps to ensure both clinical and modelling quality control, and the automation and standardisation of each step to ensure accurate comparison in clinically appropriate timeframes. Another benefit of TAVR-AID is reproducibility, as the processing steps can be saved, providing an auditable sequence of operations that can be analysed post-simulation.

[0036] The TAVR-AID pipeline goes beyond visualisation by adding an Al-based recommendation module which outputs information that is critical to the TAVR success, including the valve type, size, position, and orientation. It also provides outputs on the longevity of the valve and the risk of complications. In an embodiment, a supervised machine learning algorithm, for example XGBoost, can be used. The inputs to the algorithm are features derived from the simulation, for example one or more of the distribution of the velocity of the blood, number of vorticities, Reynold’s number relative to the valve, peak wall shear stress, and other turbulence-related flow metrics, along with other data available including measurements derived from the CT, demographic data from the patient’s EHR. These inputs are used in a classifier to predict the valve type, and regressors which predict the valve’s position and orientation along with its longevity. The classifier can be trained using a leave-one-out cross-validation technique to characterize its performance. Desirably, data for patients of different sex and different valve sizes is used for training.

[0037] In another embodiment, a deep neural network is used to make the recommendations mentioned above. This network may be a convolutional -neural network or a vision transformer. In this case, instead of inputting derived features from the simulation, the entire 4D (3D + time) simulation is input into the deep neural network, along with themeasurements derived from the CT, demographic data from the patient’s electronic healthcare record (EHR). The network then extracts features from the 4D simulation and combines them with the non-simulation features such as measurements derived from the CT, demographic data from the patient’s EHR.

[0038] Variables that may be used to predict outcomes include one or more of: age, sex, chronic conditions, smoking history, dietary habits, physical activity, ECG results, blood pressure, previous surgery. In particular, clinical measurements, e.g. gained from the ultrasound scans and / or CT image, including velocities and dimensions are useful. Such measurements may include one or more of: the aortic annulus diameter; annual shape and area; dimensions of the left ventricular outflow tract (LVOT); and velocity of the blood flow. Where a different vessel is modelled, corresponding dimensions are be used instead.

[0039] Embodiments of the invention can be used to simulate different treatments and therefore is ideal for “what if’ scenarios, for example “what if a smaller valve is used?” or “what if we move the valve to a different position?”. In this way the invention can help the cardiology team choose an optimal treatment within the set of TAVR possibilities. Further, the invention can help the cardiology team choose an optimal treatment from the broader possibilities of treatment approach. An alternative to TAVR is surgical aortic valve replacement (SAVR) - that is, open heart surgery. The invention might help decide between TAVR and SAVR, for example due to the choice of valve being used and / or the predicted longevity of TAVR. Sometimes the flow through a TAVR may outperform a SAVR and vice versa.

[0040] Valves primarily used in SAVR are mechanical valves, such as valves comprising durable materials which may be designed for long-term durability. The valves in TAVR are typically bioprosthetic valves (rather than mechanical ones), such as valves comprising animal and / or human tissue which may be mounted on a collapsible stent frame. However, bioprosthetic valves may be used in SAVR and mechanical, or other valve types, may be developed for use in TAVR. The invention might help decide between treatment options, by the assessment of the performance of one or more valves (bioprosthetic valve(s), and / or mechanical valve(s), and / or other valve(s)) and / or comparison of performance between different valves. The invention might help decide between TAVR and SAVR, for example due to the desired position of the valve and the ease or possibility of achieving (near) the desired position by TAVR and SAVR.

[0041] The invention may assist in the selection of a specific valve; in particular selection of a type of valve (e.g. balloon expandable or self-expanding) and / or details of the valve such as size, orientation and height. For balloon expandable valves, invention may assist in determining the balloon size, which can be affected by adding or subtracting contrast from the balloon, thereby providing a better fit.

[0042] The invention is also applicable to “valve-in-valve” (ViV) interventions, in particular to valve-in-valve transcatheter aortic valve replacement (ViV TAVR) in which a new valve is inserted by TAVR into the orifice of a failed replacement valve. The failed valve is not removed. ViV TAVR may include, for example, TAVR inside of a failed surgical valve (TAVR-in-SAVR), TAVR inside of a failed TAVR valve (TAVR-in-TAVR), and / or TAVR inside of a failed TAVR valve, which was previously placed in failed SAVR valve (TAVR-in-TAVR-in-SAVR).

[0043] The invention can be used to simulate treatment with different valve features (for example, stented, sutureless, stentless) and / or leaflet orientation (for example, externally or internally mounted leaflets). The invention can help the cardiology team choose a more optimal treatment from the broader possibilities of treatment approach.

[0044] The output of the invention may include visual representations of a suggested placement of an implanted device in order to assist the operator in achieving the suggested placement. The suggested valve type, size, position, and orientation may be included, for example. Such a visualisation may be used in an iterative process whereby the operator makes adjustments, e.g. to placement of the implanted device, and the effects of such adjustments are modelled. In particular, it is believed the turbulence in the vicinity of the valve reduces the lifetime of the implanted valve. Therefore, the operator may view a video of simulated turbulence and make adjustments that seek to reduce turbulence; then new simulations are performed and the operator views the new simulations to confirm whether there is an improvement. This iterative approach may be continued as long as required.

[0045] Such visualisations may be effective to explain procedures to individuals lacking medical training, e.g. the patient who may better give informed consent. The present invention can therefore present information concerning the predicted function of an implanted device, for example the effect of an implanted valve on turbulent flow in a vessel, in a manner that is objectively easier to understand.

[0046] In this way, patient outcomes are improved, the longevity of the valve is improved, and the risk of complications is reduced. This is achieved in a way that is further to and separate from outcomes that depend on the intervention process itself. That is, an intervention performed to the same (e.g. average) degree of competence based on the output of the invention (a suggested placement of a suggested implanted device) is inherently more likely to provide better outcomes than an intervention performed to the same level of competence but without the knowledge of the output of the invention. The present invention can guide the operator in a way that improves patient outcomes.

[0047] An important element of TAVR-AID is high resolution simulation of aortic blood flow. Desirably, the resolution is sufficient to be able to faithfully capture the hemodynamics, particularly around the valve. The simulation can be considered well-resolved if the resolution is high enough that a threshold amount, e.g. 80%, of the total turbulent kinetic energy is captured by the simulation. Detailed flow dynamics allows for the capture of subtle flow patterns that are critical in understanding how the implanted valve will interact with the blood flow. This includes the ability to model vorticities, flow separation, and recirculation zones with greater accuracy, which are important for assessing the risk of thrombosis, valve deterioration, paravalvular leak, and other hemodynamic-related complications. It can also provide a detailed analysis of wall shear stress (WSS) on the aorta. Elevated WSS can lead to endothelial damage, which is a precursor to atherosclerosis and other vascular diseases. Finally, we define novel measures of success based on flow patterns. The success of the procedure is defined based on analyzing the changes in flow patterns (turbulent flow) using the Reynolds number before and after the TAVR procedure. Also, TAVR-AID uses blood flow velocity to define a successful outcome: the median flow velocity is used as a parameter to determine a successful outcome in terms of flow patterns. Equating flow parameters to clinical outcomes, as TAVR-AID aims to link the flow parameters measured during the procedure with long-term clinical outcomes, providing a more immediate way to assess success compared to waiting for rare events like stroke.

[0048] TAVR-AID focuses on using high-quality data and a novel definition of success based on flow patterns, potentially leading to a more accurate and efficient way to assess TAVR procedures.

[0049] We use an Al-based CT image segmentation approach [9] to extract the aorta anatomy from a CT image. The approach utilises the popular nnUNet architecture [6] as abackbone and adds a misclassification loss designed to reduce false positive and false negative predictions, producing a more faithful representation of the aorta’s geometry. The method was trained on CT angiography images, but is also effective for segmentation of CT images throughout the cardiac cycle, including best systole and best diastole. Figure 2 provides examples of the segmentation algorithm [9] applied to held-out images not used in testing. In (a), we show a CT angiography image, with axial (upper left), coronal (lower left) and sagittal (lower right) views, along with the segmentation produced in the upper right, (b) shows a similar example for best systole. Note the systolic image captures less of the aortic vessel tree compared to the angiography image.

[0050] Next, clipping is performed in the TAVR-AID pipeline. Clipping enables the digital twin simulation to focus on the aortic root where the TAVR device is placed, and the adjacent ascending aorta region. For TAVR and / or for interventions in other vessels, clipping is performed to ensure the simulation encompasses the key anatomical areas of interest and avoids unnecessary extra regions. Clipping can reduce the computational load for simulation, thereby allowing use of less computing resource or providing results more quickly. A centreline is extracted, and clipping is performed orthogonal to the centreline. At the aortic root, the aortic leaflets are visible in the plane orthogonal to the centreline. The modeling may also include the LVOT to accurately model the blood flow in the base of the aorta. Figure 3 provides examples of the clipping applied to the segmentation. This step produces the aortic region that will be used in the digital twin simulation. In (a), visualisation of the region of interest for the digital twin is shown. In (b), the centreline (blue) and cross-sectional slices (yellow) following the centreline are shown. In (c), visualisations of the leaflets in the plane orthogonal the centreline at the aortic root are shown along with the 3D surface that is extracted by the clipping process.

[0051] Once the surface geometry is defined, the meshing operation produces a solid model comprised of tetrahedral elements required for simulation. The resolution of the tetrahedra is an adjustable parameter, with high resolution providing a more precise simulation of computational fluid dynamics, but requiring a longer time to simulate. Figure 4 provides examples of the meshing steps applied to the clipped segmentation. The image on the left represents the initial solid model produced after the clipping step. On the right, the resolution of the solid model is increased, producing smaller tetrahedral elements, and the inflow surface (purple) and outflow surface (orange) are identified along with the outersurface (blue). In addition to producing a solid model, this step also identifies specific surfaces where boundary conditions are applied, in particular the inflow surface, outflow surface, and vessel walls, shown in purple, orange, and light blue respectively in the rightmost image of Figure 4.

[0052] The inflow into the aorta is modelled as a jet. We approximate the aortic annulus as a circular opening, although other possibilities exist. Under the circularity assumption, the radius of the jet rjet can be estimated aswhere AVA is the aortic valve area. At the input boundary surface, the jet is modeled as flow and no-flow regions, as shown in Figure 5(a), with the flow region denoted as purple and the no flow region as aqua. Within the flow region, we use a parabolic velocity profile shown in (b), so that blood flows with maximum velocity in the centre and falls off from the centre, as:where uo is the maximal velocity provided by the echocardiography image. This profile is shown in Figure 5(b). The jet also varies temporally, also based on the echocardiography as shown in Figure 5(c) which shows the maximal velocity of the blood as a function of time, modeled as the blue curve fit to the echocardiographic image. A plug profile inflow velocity can also be implemented. A 0D vasculature model, such as the three-element Windkessel model, derived using the inflow condition and echocardiography data can be used as the outflow condition.

[0053] At this stage, the TAVR-AID pipeline is ready for simulating blood flow. The simulation is achieved using a solver, which applies computational fluid dynamics to model the blood flow. This produces a 3D + time simulation of the blood flow in the aorta. This hemodynamic data can be analysed, particularly for vorticities, using the vorticity magnitude co = V x u, where u is the simulated velocity of the blood. Other measures such as the Q- criterion, which defines vortices as areas where the vorticity magnitude is greater than the magnitude of the rate of strain, can be visualised as well. Examples are shown in Figure 6 which depicts simulation results, taken from a video showing the movement of blood in the aorta. In (a), a visualisation of the vorticity magnitude is shown. In (b), a visualisation of the Q criterion is shown.

[0054] The last stage of the TAVR-AID pipeline is a multi-modal artificial intelligence module, which takes the simulated blood flow as input, along with measurements from the CT and echocardiography images, and clinical data from the patient’s electronic healthcare record. The Al makes predictions about the valve type, position, orientation, and longevity, providing pre-interventional clinical decision support to the multidisciplinary heart team. A simple implementation of this Al component is a tree-based learner like XGBoost, which can take in tabular data such as the number of vorticities and their location, measurements, clinical data, and output the predictions described above. A more advanced method would rely on a conditional neural network architecture, where the 4D (3D+time) data is input to the model, with features modulated by the tabular data similar to FiLM-UNet

[0012] ,

[0055] TAVR-AID, is a digital twin pipeline that provides mechanistic simulation of a patient’s blood flow coupled with predictive capabilities of artificial intelligence. TAVR-AID is used as a pre-interventional decision support tool to optimise the TAVR procedure, including the choice of the prosthetic valve, its position and orientation, along with predictions about the longevity of the valve.

[0056] As well as supporting medical procedures, the techniques described herein can be used to support implant design. In any of these methods, the implant may be a prosthetic cardiac valve. In particular, there is provided a method for supporting implant design, the method comprising: obtaining a three-dimensional image of a vessel of an individual; deriving measurements of the vessel from the three-dimensional image; constructing a model of fluid flow in the vessel using the measurements; and using a trained machine learning model to output one or more recommendations, based on outputs of the model of fluid flow, for design of an implant to be inserted into the vessel. In another example, there is provided a method for supporting implant design, the method comprising: obtaining a three-dimensional image of a vessel of an individual; obtaining a first implant design of an implant to be inserted into the vessel; deriving measurements of the vessel from the three-dimensional image; constructing a model of fluid flow in the vessel using the measurements and the first implant design, such that the first implant design influences the fluid flow in the model; obtaining a modified implant design; constructing an updated model of fluid flow in the vessel using the measurements and the modified implant design, such that the modified implant design influences the fluid flow in the updated model. In these methods, any of the features of methods already described herein may be applied, in any combination.

[0057] Methods for supporting implant design may comprise obtaining an implant design, and constructing the model of fluid flow in the vessel based on the implant design (as well as the measurements of the vessel), such that the implant design influences the fluid flow in the model. The implant design may comprise a three-dimensional mesh, which may be derived from a computer-aided design (CAD) drawing.

[0058] These methods can be used to compare candidate designs for new implants. For example, two or more candidate implant designs may be compared. The fluid flow may be modelled for each of the candidate implant designs. The model may output a comparison between the candidate implant designs, for example, the model may identify the candidate implant design with a better predicted success for implantation (for example using any of the success parameters discussed herein). The model may be constructed for a plurality of vessel geometries from a plurality of individuals, such that the identification of the model with the best predicted success rate is based on the simulated performance of the candidate implants across a population.

[0059] Additionally or alternatively to selecting from a set of candidate models, the recommendation output by the machine learning model may comprise an indication of an adjustment to be made to a structural parameter of the implant design. This recommendation may indicate a specific structural parameter and specify how to adjust the parameter. For example, the parameter could be a material property of the valve, such as a material stiffness parameter which is relevant to turbulence and may affect the flutter of the valve. In this case, the model may recommend increasing or decreasing the material stiffness parameter, and may recommend the amount by which the material stiffness parameter be increased or decreased. It will be appreciated that the model may recommend adjusting different parameters, properties or dimensions of the valve. The model may recommend how to adjust the parameter (e.g. increase, decrease) and / or the degree to which the parameter / property / dimension should be adjusted (e.g. a number of units of measurement of the relevant parameter / property / dimension).

[0060] Additionally or alternatively, the recommendation could comprise identifying one or more features of the fluid flow in the model and recommending that the implant design be adjusted to influence the one or more features. For example, the machine learning model might identify regions of turbulence that are undesirable, so that a human may consider how to adjust the design to address the predicted turbulence.

[0061] Additionally or alternatively, the machine learning model may generate a new implant design based on the simulated flow for an individual, or for a group of individuals. The trained machine learning model may be a generative deep learning model. The machine learning model may have been trained on simulations of a range of valve designs for a population of individuals.

[0062] Any of the methods of implant design discussed herein may further include outputting a design document for the implant according to the one or more recommendations produced by the trained machine learning model. The methods may further comprise making an implant according to recommendations and / or the design document.Conclusion

[0063] The methods of the present invention may be performed by computer systems comprising one or more computers. A computer used to implement the invention may comprise one or more processors, including general purpose CPUs, graphical processing units (GPUs), tensor processing units (TPU) or other specialised processors. A computer used to implement the invention may be physical or virtual. A computer used to implement the invention may be a server, a client or a workstation. Multiple computers used to implement the invention may be distributed and interconnected via a network such as a local area network (LAN) or wide area network (WAN). Individual steps of the method may be carried out by a computer system but not necessarily the same computer system. Results of a method of the invention may be displayed to a user or stored in any suitable storage medium. The present invention may be embodied in a non-transitory computer-readable storage medium that stores instructions to carry out a method of the invention. The present invention may be embodied in a computer system comprising one or more processors and memory or storage storing instructions to carry out a method of the invention. The present invention may be incorporated into a medical imaging device or into software updates or add-ons for such a device.

[0064] Having described the invention it will be appreciated that variations may be made on the above described embodiments which are not intended to be limiting. The invention has been described in relation to human subjects. It will be appreciated that the invention may also be applied to other animals. The invention is defined in the appended claims and their equivalents.

[0065] References[1] Shelly Singh-Gryzbon amd Beatrice Ncho, Vahid Sadri, Sanchita S. Bhat, Sai S. Kollapaneni, Dharani Balakumar, Zhenglun A. Wei, Philipp Ruile, Franz -Joseph Neumann, Philipp Blanke, and Ajit P. Yoganathan. Influence of patient specific characteristics on transcatheter heart valve neo-sinus flow: An in silico study. Annals of Biomedical Engineering, 48, 2020.[2] Matteo Bianchi, Gil Marom, Ram P Ghosh, Oren M Rotman, Puja Parikh, Luis Gruberg, and Danny Bluestein. Patient-specific simulation of transcatheter aortic valve replacement: impact of deployment options on paravalvular leakage. Biomechanics and Modeling in Mechanobiology, 18, 2019.[3] Cameron Dowling, Robert Gooley, Liam McCormick, Rahul P. Sharma, Alan C. Yeung, William F. Fearon, James Dargan, Faisal Khan, Sami Firoozi, and Stephen J. Brecker. Ongoing experience with patient-specific computer simulation of transcatheter aortic valve replacement in bicuspid aortic valve. Cardiovascular Revascularization Medicine, 51, 2023.[4] Ivan Fumagalli, Rebecca Polidori, Francesca Renzi, Laura Fusini, Alfio Quarteroni, Gianluca Pontone, and Christian Vergara. Fluid-structure interaction analysis of transcatheter aortic valve implantation. International Journal for Numerical Methods in Biomedical Engineering, 39, 2023.[5] Hoda Hatoum, Shelly Singh-Gryzbon, Fateme Esmailie, Philipp Ruile, Franz -Josef Neumann, Philipp Blanke abd Vinod H. Thourani, Ajit P. Yoganathan, and Lakshmi Prasad Dasi. Predictive model for thrombus formation after transcatheter valve replacement. Cardiovascular Engineering and Technology, 2021.[6] Fabian Isensee, Paul F. Jaeger, Simon A. A. Kohl, Jens Petersen, and Maier-Hein Klaus H. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18, pages203-211, 2021.[7] Karim Kadry, Shreya Gupta, Farhard R. Nezami, and Elazer R. Edelman. Probing the limits and capabilities of diffusion models for the anatomic editing of digital twins. arXiV, 2023.[8] Harkamaljot S Kandail, Setu D Trivedi, Armaan C Shaikh, Tanvir K Bajwa, Daniel P O’Hair, Arshad Jahangir, and John F LaDisa. Impact of annular and supra-annular corevalvedeployment locations on aortic and coronary artery hemodynamics. Journal of the Mechanical Behavior of Biomedical Materials, 86: 131-142, 2018.[9] Abbas Khan, Muhammad Asad, Alexander Zolotarev, Caroline Roney, Anthony Mathur, Martin Benning, and Gregory Slabaugh. Misclassification loss for segmentation of the aortic vessel tree. In Segmentation of the Aorta. Towards the Automatic Segmentation, Modeling, and Meshing of the Aortic Vessel Tree from Multicenter Acquisition: First Challenge, SEG.A. 2023, 2023.

[0010] Seyedvahid Khodaei, Louis Garber, Mohamed Abdelkhalek, Nima Maftoon, Ali Emadi, and Zahra Keshavarz-Motamed. Reducing long-term mortality post transcatheter aortic valve replacement requires systemic differentiation of patient specific coronary hemodynamics. Journal of the American Heart Association, 12, 2023.

[0011] Alex G. Kuchumov, Anastasiya Makashova, Sergey Vladimirov, Vsevolod Borodin, and Anna Dokuchaeva. Fluid-structure interaction aortic valve surgery simulation: A review. Cardiovascular Revascularization Medicine, 8(11) 295, 2023.

[0012] Andreanne Lemay, Charley Gros, Olivier Vincent, Yaou Liu, Joseph Paul Cohen, and Julien Cohen- Adad. Benefits of linear conditioning with metadata for image segmentation. In Medical Imaging with Deep Learning, 2021.

[0013] Daniel H Pak, Minliang Liu, Caglar Kim, Theodore Ozturk, Raymond McKay, Ellen T. Roche, Rudolph Gleason, and James S. Duncan. Robust automated calcification meshing for biomechanical cardiac digital twins. ArXiv preprint, 2024.

[0014] Daniel H Pak, Minliang Liu, Theodore Kim, Liang Liang, Andres Caballero, John Onofrey, Shawn S Ahn, Yilin Xu, Raymond McKay, Wei Sun, Rudolph Gleason, and James S. Duncan. Patient-specific heart geometry modeling for solid biomechanics using deep learning. IEEE Transactions on Medical Imaging, 2024.

[0015] Leonardo Pietrasanta, Shaokai Zheng, Dario De Marinis, David Hasler, and Dominik Obrist. Characterization of turbulent flow behind a transcatheter aortic valve in different implantation positions. Frontiers in Cardiovascular Medicine, 8, 2022.

[0016] Amir Rouhollahi, James Noel Willi, Sandra Haltmeier, Alireza Mehrtash, Ross Straughan, Hoda Javadikasgari, Jonathan Brown, Akinobu Itoh, and Kim I. de la Cruz. Cardiovision: A fully automated deep learning package for medical image segmentation and reconstruction generating digital twins for patients with aortic stenosis. Computerized Medical Imaging and Graphics, 109: 102289, 2023.

[0017] Anas M. Tahir, Onur Mutlu, Faycal Bensaali, Rabab Ward, Abdel Naser Ghareeb, Sherif M. H. A. Helmy, Khaled T. Othman, Mohammed A. Al -Hashemi, Salem Abujalala, Muhammad E. H. Chowdhury, A.Rahman D. M. H. Alnabti, and Huseyin C. Yalcin. Latest developments in adapting deep learning for assessing tavr procedures and outcomes. Journal of Clinical Medicine, 2023.

[0018] Stefan Toggweiler, Moritz C. Wyler von Ballmoos, Federico Moccetti, Andre Douvemy, Mathias Wolfrum, Ziya Imamoglu, Anton Mohler, Utku Giilan, and Won-Keun Kim. A fully automated artificial intelligence-driven software for planning of transcatheter aortic valve replacement. Cardiovascular Revascularization Medicine, 2024.

[0019] Frank M. Weber, Thomas Stehle, Irina Waechter-Stehle, Michael Gbtz, Jochen Peters,Sabine Mollus, Jan Balzer, Malte Keim, and Juergen Weese. Analysis of mitral valve motion in 4d transesophageal echocardiography for transcatheter aortic valve implantation. Statistical Atlases and Computational Models of the Heart - Imaging and Modelling Challenges, 2015.

Claims

CLAIMS1. A method for supporting a medical procedure, the method comprising: obtaining a three-dimensional image of a vessel of a patient; deriving measurements of the vessel from the three-dimensional image; constructing a model of fluid flow in the vessel using the measurements; and using a trained machine learning model to make predictions about the medical procedure based on outputs of the model of fluid flow.

2. A method according to claim 1 wherein the medical procedure comprises implantation of a device and the predictions comprise at least one of: a recommendation of a size or type of device; a recommendation of a placement of the device; and a predicted lifetime of the device in situ.

3. A method according to claim 1 or 2 wherein the medical procedure comprises an Aortic Valve Replacement, desirably a Transcatheter Aortic Valve Replacement, a TAVR valve-in- valve procedure or a valve-in-valve procedure.

4. A method according to any one of the preceding claims further comprising obtaining ultrasound measurements of fluid flow in the vessel prior to the medical procedure and wherein the measurements of fluid flow are used to provide boundary conditions for the model of fluid flow.

5. A method according to any one of the preceding claims wherein constructing a model of fluid flow comprises segmentation of the three-dimensional image to obtain a three- dimensional mesh of vessels and clipping of the three-dimensional mesh to remove unnecessary parts of the vessel.

6. A method according to claim 5 when dependent on claim 4 wherein clipping of the three-dimensional mesh comprises clipping the three-dimensional mesh at a location at which a measurement of fluid flow was obtained.

7. A method according to any one of the preceding claims wherein the trained machine learning model is a supervised-learning classifier.

8. A method according to claim 7 wherein outputs of the model of fluid flow input to the trained machine learning model are features derived from the simulation, for example one or more of: the distribution of the velocity of the blood; number of vorticities; Reynold’s number relative to the valve; peak wall shear stress.

9. A method according to any one of claims 1 to 6 wherein the trained machine learning model is a deep neural network.

10. A method according to claim 9 wherein outputs of the model of fluid flow input to the trained machine learning model comprise a four-dimensional simulation of fluid flow in the vessel.

11. A method according to any one of the preceding claims wherein the predictions of the trained machine learning model are additionally based on other data including at least one of: measurements derived from the three-dimensional image; and demographic data of the patient.

12. A method according to any one of the preceding claims wherein the three-dimensional image is obtained by computed tomography.

13. A method according to any one of the preceding claims wherein the method comprises: recommendation of a transcatheter heart valve; and performing at least part of the interventional procedure, wherein the interventional procedure comprises implantation of the transcatheter heart valve.

14. A method according to any one of the preceding claims wherein at least part of the interventional procedure is performed under image-guidance.

15. A method according to any one of the preceding claims wherein at least part of the interventional procedure is performed without direct visualisation of an implantation site.

16. A method according to any one of the preceding claims wherein the at least part of the interventional procedure is an open surgical procedure that comprises direct visualisation of an implantation site.

17. A method according to any one of the preceding claims wherein the method comprises: recommendation of a surgical heart valve; and performing at least part of the interventional procedure, wherein the interventional procedure comprises surgical implantation of the heart valve.

18. A method according to any one of the preceding claims wherein the method comprises: recommendation of a stent device; and performing at least part of the interventional procedure, wherein the interventional procedure comprises implantation of the stent device.

19. A method for supporting implant design, the method comprising: obtaining a three-dimensional image of a vessel of an individual; obtaining a first implant design of an implant to be inserted into the vessel; deriving measurements of the vessel from the three-dimensional image; constructing a model of fluid flow in the vessel using the measurements and the first implant design, such that the first implant design influences the fluid flow in the model; obtaining a modified implant design; constructing an updated model of fluid flow in the vessel using the measurements and the modified implant design, such that the modified implant design influences the fluid flow in the updated model.

20. The method of claim 19, further comprising using a trained machine-learning model to output one or more recommendations, based on outputs of the model of fluid flow and outputs of the updated model of fluid flow, for design of the implant,wherein, optionally, the recommendations comprise a comparison between the first and modified implant designs, and, further optionally, the comparison comprises identifying the one of the first implant design and the modified implant design with a better predicted success for implantation.

21. A method for supporting implant design, the method comprising: obtaining a three-dimensional image of a vessel of an individual; deriving measurements of the vessel from the three-dimensional image; constructing a model of fluid flow in the vessel using the measurements; and using a trained machine learning model to output one or more recommendations, based on outputs of the model of fluid flow, for design of an implant to be inserted into the vessel.

22. The method of claim 21, further comprising obtaining an implant design, wherein constructing the model of fluid flow in the vessel is further based on the implant design, such that the implant design influences the fluid flow in the model.

23. The method of claim 19, 20 or 22, wherein the implant design comprises a three- dimensional mesh.

24. The method of claim 22 or 23, wherein the implant design is a first candidate implant design, the method further comprises obtaining a second candidate implant design and constructing a model of fluid flow in the vessel using the measurements and based on the second candidate implant design, and the recommendations comprise a comparison between the first and second candidate implant designs.

25. The method of claim 24, wherein the comparison comprises identifying the candidate implant design with a better predicted success for implantation.

26. The method of any one of claims 20 to 25, wherein the steps of obtaining a three- dimensional image of a vessel of an individual, deriving measurements of the vessel from the three-dimensional image, and constructing a model of fluid flow in the vessel using the measurements, are performed for a plurality of individuals, and the recommendations are based on outputs of the models of fluid flow for the plurality of individuals.

27. The method of any one of claims 20 and 22 to 26 , wherein the recommendation comprises an indication of an adjustment to be made to a structural parameter of the implant design.

28. The method of any one of claims 20 and 22 to 27, wherein the recommendation comprises identifying one or more features of the fluid flow in the model and recommending that the implant design be adjusted to influence the one or more features.

29. The method of any one of claims 20 to 28, wherein the recommendation includes a new implant design generated by the trained machine learning model.

30. The method of any one of claims 19 to 29, wherein the implant is a prosthetic cardiac valve.

31. The method of any one of claims 19 to 30, further comprising outputting a design document for the implant according to the one or more recommendations and, optionally, making an implant according to the design document.

32. The method of any one of claims 19 to 31, further comprising making an implant according to the one or more recommendations.

33. A method according to any one of claims 19 to 32, further comprising obtaining ultrasound measurements of fluid flow in the vessel and wherein the measurements of fluid flow are used to provide boundary conditions for the model of fluid flow.

34. A method according to any one of claims 19 to 33, wherein constructing a model of fluid flow comprises segmentation of the three-dimensional image to obtain a three- dimensional mesh of vessels and clipping of the three-dimensional mesh to remove unnecessary parts of the vessel.

35. A method according to claim 34 when dependent on claim 33, wherein clipping of the three-dimensional mesh comprises clipping the three-dimensional mesh at a location at which a measurement of fluid flow was obtained.

36. A method according to any one of claims 19 to 35, wherein the trained machine learning model is a supervised-learning classifier.

37. A method according to claim 36, wherein outputs of the model of fluid flow input to the trained machine learning model are features derived from the simulation, for example one or more of: the distribution of the velocity of the blood; number of vorticities; Reynold’s number relative to the valve; peak wall shear stress.

38. A method according to any one of claims 19 to 37, wherein the trained machine learning model is a deep neural network.

39. A method according to claim 38, wherein outputs of the model of fluid flow input to the trained machine learning model comprise a four-dimensional simulation of fluid flow in the vessel.

40. A method according to any one of claims 19 to 39, wherein the predictions of the trained machine learning model are additionally based on other data including at least one of: measurements derived from the three-dimensional image; and demographic data of the individual.

41. A method according to any one of claims 19 to 40, wherein the three-dimensional image is obtained by computed tomography.

Images